Graph Neural Network Encoding for Community Detection in Attribute Networks

Area of Science:

• Graph Neural Networks
• Network Science
• Artificial Intelligence

Background:

• Community detection in complex attribute networks is a challenging problem.
• Existing methods often struggle with the complexity and attributes of large networks.
• Multiobjective evolutionary algorithms (MOEA) offer a framework for optimization but require effective encoding strategies.

Purpose of the Study:

• To propose a novel graph neural network (GNN) encoding method for MOEA in attribute network community detection.
• To develop new objective functions for evaluating community attribute homogeneity in single and multi-attribute networks.
• To introduce a continuous encoding MOEA (CE-MOEA) for solving the transformed community detection problem.

Main Methods:

• A GNN encoding method associates continuous variables with network edges, transformed into discrete community solutions.
• Two objective functions are proposed to measure node attribute homogeneity within communities.
• A MOEA based on NSGA-II, termed CE-MOEA, is developed using the GNN encoding and objective functions.

Main Results:

• Experimental results demonstrate that the CE-MOEA significantly outperforms existing evolutionary and non-evolutionary algorithms on various attribute networks.
• The GNN encoding transforms community detection problems into smoother fitness landscapes compared to original formulations.
• The proposed method shows superior performance in identifying communities with high attribute homogeneity.

Conclusions:

• The proposed GNN encoding method effectively addresses the community detection problem in complex attribute networks.
• CE-MOEA provides a robust and efficient approach for attribute network analysis.
• The findings highlight the potential of GNNs in enhancing evolutionary algorithms for network science applications.